Figure 1.

• A. This figure is based on the average expression measurements of all genes. They often have a bi-modal (two-peak) distribution: a high and narrow peak on the left corresponding to inactive genes and a relatively lower and wider peak on the right corresponding to active genes expressed different levels.
• B. This figure is based on the between-sample standard deviations of all genes. The distribution usually has a single peak and a long tail on the right side. An enlarged tail usually indicates more genes are differentially expressed between samples. In a high quality data set, a gene not differentially expressed between samples should have small variance. So, if we assume that the majority of genes have no or little differential expression, the skewness of the peak towards the left side roughly represents data quality.
• C. This figure summarizes the relationship between gene expression measurements and their between-sample variance. If there is a strong dependency of variance on expression measurements, the analysis of differential expression will be affected by the baseline expression of genes. Variance stabilization procedure is commonly used during data normalization to minimize such dependency. In the figure, the red line is the result of LOWESS smoothing and the majority of this line should be approximately straight and horizontal if the data has been normalized properly.

Figure 2. Each boxplot summarizes the expression measurements of all genes in one sample. Based on the assumption that all samples have approximately the same global distribution of gene expression measurements, all boxes should look similar regardless of which group they belong to. This assumption is not always true though. For example, normal cells and cancer cells could have dramatically different global patterns of gene expression. The notches of each box indicate the median, the lower and upper sides of the box represent the first and third quantile, and the individual data points out of the whiskers are the outliers (more than 1.5 inter-quantile range from the median).

Figure 3. This is an unsupervised clustering of samples used all genes in the data set. On the clustering tree, the vertical location of their lowest common node of any two samples represents their similarity (lower = more similar). Since the sample grouping information is not used, the splitting of samples into two sub-trees will indicate that these samples belong to different groups due to a known or unknown factor. Unexpected splitting often suggest outliers or confounding variables.

Figure 4. Principal Components Analysis (PCA) is also an unsupervised analysis that converts a large number of correlated variables (genes) into a smaller set of uncorrelated variables called principal components (PCs). Each principal component accounts for certain percentage of total variability of a data set so the PCs can be ordered by their percentages. This figure plots the top two PCs on the two axes. In general, samples closer to each other have more similar gene expression patterns. PCA can be used to identify sample features, such as age, disease, and treatment, that are associated with one or two PCs. It then can be concluded that these features are responsible for part of the total variability in the data set.

Figure 5. The differential expression of all 10162 genes can be visualized in different ways:

• A: M-A Plot. This plot visualizes global pattern of group difference (as log2-ratio). M (Y axis) is the log2-ratio and A (X axis) is the average measurements of the two groups. Each dot represents a gene and the green line is generated by LOWESS smoothing. The location and shape of the LOWESS line indicate whether there is an overall skewness of differential expression or expression level-dependent differential expression.
• B: Volcano Plot. This plot puts together both types of variables (usually p value and log2-ratio) used to represent differential expression and their association. A common strategy is to only consider genes with both small p values and large log2-ratios as GOI (genes of interest). Each dot represents a gene and its size is proportional to its distance from the [0, 1] point. The horizontal line corresponds to a p value of 0.05 and the vertical lines correspond to 2.0 fold change on both sides.
• C: P Value Distribution. This plot shows the counts of genes whose Rank Product p values are located in each 0.01 interval. So, the leftmost bar indicates the number of genes having p values less than 0.01. If the data is completely random and there are enough samples, the p values will have a uniform distribution, with approximately the same number of genes in each interval. If there are indeed a number of genes differentially expressed between the compared groups and the microarray experiments are properly designed and executed, we expect that the p values have a distribution with the highest density on the left side. If the highest density of the distribution is in the middle or on the right side, we will suspect a confounding factor that has a bigger influence on the data than the factor distinguishing the two compared groups.
• D: FDR. This plots traces the number of genes corresponding to each FDR (false discovery rate) cutoff. Table 1 lists the number of genes corresponding to commonly used FDR cutoffs.

Table 1. Number of top DEGs selected via different cutoffs of FDR. FDRs are calculated using the Benjamini&Hochberg method (Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society Series B 57, 289–300. 1995)

Figure 6. Top-ranked genes with increased (left) and decreased (right) expression in Patient. Click here to view differential expression of all genes.

Figure 7. Heatmap of the top 100 and 100 DEGs with higher (red) and lower (yellow) expression in Patient. Each row represents a DEG, whose expression measurements are normalized across samples. Samples are clustered by these genes and the columns are colored (blue = Control and red = Patient).